- Download Places365 to the
Dataset
directory. - Extract
places365_train_standard.txt
fromplaces365_train_standard.zip
to theDataset
directory. - Start training the model by running all the blocks in
UNet_PatchGan_v4.ipnyb
. - To evaluate the model with new pictures, modify
ckpt_path
andimg_path
in the third and last blocks ofEvaluate.ipnyb
respectively.
To change the root directory of datasets, modify dataset.py
on Line 8 and 34. Be sure to include places365_train_standard.txt
in the new directory.
- The model is mainly based on PatchGan. A classifier and Convolution Block Attention Module (CBAM) are incorporated to improve the model's performance.
- For this model, pictures are represented in LAB color space. There are two main reasons why we utilize LAB instead of RGB: Given that the "L" channel can be used as input, the model only needs to generate and concatenate values in the "A" and "B" channels; on the other hand, it will be easier for the model to colorize images of varying size, since the rescaled "A" and "B" channels can be concatenated using the original "L" channel with no distortion in pictures.